Impact of Surgical Resection on 5‐Year Survival in Elderly
NSCLC
Patients: A
SEER
Database Analysis With Machine Learning‐Based Predictive Modeling
Bo Zhou, Xiangchun Xu, Baowei Ma, Huachuan Wang ABSTRACT
Background
Surgical resection is the primary curative option for elderly patients with non‐small cell lung cancer (NSCLC), but its survival benefit remains debated due to comorbidities and competing risks. This study evaluated the dynamic impact of surgery on long‐term survival and developed machine learning models for prognostic prediction.
Methods
We analyzed 259,701 elderly NSCLC patients (≥ 60 years) from the SEER database (2000–2021). Patients with missing disease stage or tumor size were strictly excluded. Propensity score matching (1:1, caliper = 0.05), comprehensively adjusting for age, sex, stage, tumor size, chemotherapy, and radiotherapy to produce a highly balanced cohort of 15,946 patients (7973 per group). Multivariable Cox regression, piecewise time‐dependent Cox models, and stage‐stratified Kaplan–Meier curves were used to assess surgical effects. Five machine learning algorithms were trained to predict 5‐year overall survival.
Results
Surgery was independently associated with a significant overall survival benefit (adjusted HR = 0.45, p < 0.001). Time‐dependent piecewise Cox analysis revealed this protective effect was most profound within the first 2 years post‐diagnosis (HR = 0.408, p < 0.001) and progressively attenuated over long‐term follow‐up due to increasing competing risks. While a statistical survival advantage was observed across all stages, advanced‐stage results are likely confounded by selection bias. Among machine learning models predicting 5‐year survival, Logistic Regression achieved the highest discriminative performance (AUC = 0.724), closely followed by Random Forest (AUC = 0.719) and XGBoost (AUC = 0.700). SHAP analysis confirmed surgical resection, baseline stage, and tumor size as the most critical prognostic determinants.
Conclusion
Surgical resection was significantly associated with a profound but time‐attenuating survival benefit in elderly NSCLC patients. However, the observed statistical benefit in advanced stages must be interpreted with extreme caution due to unmeasured selection bias and should not replace individualized multidisciplinary evaluations. Logistic Regression and tree‐based models provide robust 5‐year survival predictions, highlighting the value of machine learning for individualized risk stratification in geriatric oncology.